Cut a 6-hour training recording into highlights

A training team used Claude Code with video-editing skills to transcribe a 6-hour live recording locally, find pauses and irrelevant parts, and build a Kdenlive project for manual touch-ups. · Nov 4, 2026

Done withCCClaude Code

What they did

Feng Yu added two skills to Claude Code: videocut-skills and cli-anything-kdenlive. He asked Claude to combine them without the paid Volcano Engine transcription API. Claude designed a local pipeline. It used faster-whisper to transcribe, applied the skill's rules to cut verbal slips and pauses, and wrote a Kdenlive project file. Feng fixed a bug in the Kdenlive CLI so the file would open. He then polished the cut by hand.

How it went

The 6-hour audio was transcribed and gave a rough cut, so the job dropped from about 3 days to a bit over 1. Many verbal slips and redundancies were left in, and the project had too many clips.

Worth knowing

The machine couldn't run faster-whisper large-v3. He used the small model at int8 and transcribed audio chunks in parallel. That cost some accuracy, which made the cut timings less precise.

Source: dteam.top · Nov 4, 2026 · Submitted by from search gap: 视频